Environment International 66 (2014) 165–173
Contents lists available at ScienceDirect
Environment International journal homepage: www.elsevier.com/locate/envint
Health effects of ambient air pollution: Do different methods for estimating exposure lead to different results? Yann Sellier a,b, Julien Galineau c, Agnes Hulin d, Fabrice Caini d, Nathalie Marquis c, Vladislav Navel d, Sebastien Bottagisi a,b, Lise Giorgis-Allemand a,b, Claire Jacquier c, Remy Slama a,b, Johanna Lepeule a,b,e,⁎, the EDEN mother–child cohort study group a
Inserm, U823, Institut Albert Bonniot, Team of Environmental Epidemiology Applied to Reproduction and Respiratory Health, Grenoble, France Université Joseph Fourier, Grenoble, France c Air Lorraine, Nancy, France d ATMO Poitou-Charentes, La Rochelle, France e Exposure, Epidemiology, and Risk Program, Department of Environmental Health, Harvard School of Public Health, Boston, MA 02115, USA b
a r t i c l e
i n f o
Article history: Received 27 June 2013 Accepted 4 February 2014 Available online 2 March 2014 Keywords: Exposure modeling Measurement error Spatial resolution
a b s t r a c t Background: Spatially resolved exposure models are increasingly used in epidemiology. We previously reported that, although exhibiting a moderate correlation, pregnancy nitrogen dioxide (NO2) levels estimated by the nearest air quality monitoring station (AQMS) model and a geostatistical model, showed similar associations with infant birth weight. Objectives: We extended this study by comparing a total of four exposure models, including two highly spatially resolved models: a land-use regression (LUR) model and a dispersion model. Comparisons were made in terms of predicted NO2 and particle (aerodynamic diameter b 10 μm, PM10) exposure and adjusted association with birth weight. Methods: The four exposure models were implemented in two French metropolitan areas where 1026 pregnant women were followed as part of the EDEN mother–child cohort. Results: Correlations between model predictions were high (≥0.70), except for NO2 between the AQMS and both the LUR (r = 0.54) and dispersion models (r = 0.63). Spatial variations as estimated by the AQMS model were greater for NO2 (95%) than for PM10 (22%). The direction of effect estimates of NO2 on birth weight varied according to the exposure model, while PM10 effect estimates were more consistent across exposure models. Conclusions: For PM10, highly spatially resolved exposure model agreed with the poor spatial resolution AQMS model in terms of estimated pollutant levels and health effects. For more spatially heterogeneous pollutants like NO2, although predicted levels from spatially resolved models (all but AQMS) agreed with each other, our results suggest that some may disagree with each other as well as with the AQMS regarding the direction of the estimated health effects. © 2014 Elsevier Ltd. All rights reserved.
1. Introduction Great advances have been achieved over the past 5 years in improving the spatial resolution of air pollution exposure models used for studying its mid- and long-term health effects. Epidemiological studies have moved from an approach based on air-quality monitors, which provides temporally resolved estimates but is unlikely to capture the within-city spatial heterogeneity of air pollutant concentrations (Lebret et al., 2000), to more spatially resolved models based either on measured concentrations, combined with geographical information ⁎ Corresponding author at: Equipe 12, Centre de Recherche INSERM-UJF U823, Institut Albert Bonniot, UJF Site Santé, BP 170 La Tronche, 38042 GRENOBLE Cedex 9, France. Tel.: +33 476 549 466. E-mail address: [email protected] (J. Lepeule).
http://dx.doi.org/10.1016/j.envint.2014.02.001 0160-4120/© 2014 Elsevier Ltd. All rights reserved.
system data (land-use regression) and/or geostatistical techniques, or on dispersion modeling (Briggs, 2005). A few studies have compared the performance of such exposure models and concluded to the existence of substantial differences (Cyrys et al., 2005; Gulliver et al., 2011; Marshall et al., 2008; Rosenlund et al., 2008). For endpoints such as mortality and respiratory health, measurement error resulting from the use of poorly spatially resolved models may have a large impact on the exposure–response relationship (Jerrett et al., 2005; Pouliou et al., 2008). Although fine spatial scale land-use regression and dispersion models are increasingly used in reproductive epidemiology (Aguilera et al., 2010; Pereira et al., 2011; Rahmalia et al., 2012; Slama et al., 2007), few systematic comparisons have been undertaken in terms of exposure estimates (Marshall et al., 2008) or health effect estimates (Brauer et al., 2008; Lepeule et al., 2010; Wu et al., 2011). Since implementing such fine scale exposure models is costly, and since air
166
Y. Sellier et al. / Environment International 66 (2014) 165–173
quality monitoring data are widely available, there is a critical need for understanding the implications of the use of such exposure models. The well-characterized EDEN mother–child cohort provides unique opportunities to study up to four exposure models to nitrogen dioxide (NO2) and particles with aerodynamic diameter less than 10 μm (PM10) in two metropolitan areas: the nearest air quality monitoring station (AQMS) model, a temporally-adjusted geostatistical (TAG) model, a land-use regression (LUR) model and a dispersion model. We previously showed that NO2 levels during pregnancy estimated by the nearest AQMS and by the TAG were moderately correlated; nevertheless they exhibited similar associations with infant birth weight (Lepeule et al., 2010). In the present study, we extended our work by considering the LUR and dispersion models, aiming at capturing fine
spatial contrasts in air pollutant concentrations and by considering PM10 estimated by the nearest AQMS and dispersion models. We compared these four NO2 exposure models and two PM10 exposure models in terms of exposure estimates during pregnancy and we investigated how the relationship with birth weight varied according to the exposure model used. 2. Methods 2.1. Study population and data collection The EDEN mother–child cohort consists in 2002 women enrolled before 26 gestational weeks at maternity wards of Nancy and Poitiers
Table 1 Characteristics of the study population, EDEN mother–child cohort. Characteristic
NO2 study area (n = 776)
PM10 study area (n = 1026)
Birth weight, g, mean (sd) Sex of the offspring, n (%) Male Female Gestational duration (weeks), n (%) 30–36 37–38 39–40 ≥41 Birth order, n (%) First Second Third or more Missing Month of conception of the child, n (%) January–March April–June July–September October–December Maternal age at conception (years), n (%) b25 25–29 30–34 ≥35 Maternal height (cm), n (%) b160 160–169 ≥170 Missing Maternal pre-pregnancy weight (kg), n (%) b50 50–59 60–69 70–79 ≥80 Missing Center, n (%) Poitiers Nancy Urbanization, n (%) Urban Suburban Rural Maternal age at end of education (years), n (%) ≤16 17–18 19–20 21–22 23–24 ≥25 Maternal active smoking (2nd trimester), n (%) No Yes Missing Maternal passive smoking (2nd trimester), n (%) No Yes Missing
3284 (512)
3278 (505)
395 (50.9) 381 (49.1)
531 (51.7) 495 (48.3)
47 (6.1) 128 (16.5) 440 (56.7) 161 (20.7)
60 (5.9) 176 (17.1) 580 (56.5) 210 (20.5)
367 (47.3) 263 (33.9) 145 (18.7) 1 (0.1)
477 (46.5) 365 (35.6) 182 (17.7) 2 (0.2)
167 (21.5) 184 (23.7) 226 (29.1) 199 (25.7)
216 (21.1) 233 (22.7) 302 (29.4) 275 (26.8)
187 (24.1) 289 (37.2) 203 (26.2) 97 (12.5)
231 (22.5) 380 (37.0) 290 (28.3) 125 (12.2)
188 (24.2) 460 (59.3) 121 (15.6) 7 (0.9)
255 (24.9) 607 (59.1) 153 (14.9) 11 (1.1)
83 (10.7) 333 (42.9) 211 (27.2) 87 (11.2) 60 (7.7) 2 (0.3)
112 (10.9) 433 (42.2) 282 (27.5) 120 (11.7) 76 (7.4) 3 (0.3)
316 (40.7) 460 (59.3)
436 (42.5) 590 (57.5)
414 (53.4) 320 (41.2) 42 (5.4)
424 (41.3) 377 (37.0) 225 (21.9)
52 (6.7) 104 (13.4) 124 (16.0) 165 (21.3) 174 (22.4) 157 (20.2)
59 (5.7) 128 (12.5) 166 (16.2) 231 (22.5) 234 (22.8) 208 (20.3)
507 (65.3) 264 (34.1) 5 (0.6)
703 (68.4) 317 (31.0) 6 (0.6)
641 (82.6) 133 (17.1) 2 (0.3)
845 (82.3) 178 (17.4) 3 (0.3)
Y. Sellier et al. / Environment International 66 (2014) 165–173
university hospitals, France (Slama et al., 2009). Conception dates ranged from 2002 to 2005 in both cities. Offspring birth weight was obtained from maternity records. Personal data (Table 1) were collected by questionnaires during pregnancy and after birth. We excluded women who did not speak or write French, had personal history of diabetes, multiple pregnancy, or intention to deliver outside the university hospital or to move out of the study area within 3 years. The study was approved by all relevant ethical committees (Comité Consultatif pour la Protection des Personnes dans la Recherche Biomédicale, Le Kremlin-Bicêtre University Hospital, and Commission Nationale de l'Informatique et des Libertés), and all participating women gave informed written consent for their own participation and that of their children. 2.2. Exposure assessment to air pollutants and study area Exact home addresses of the women during pregnancy were collected and geocoded. We considered four exposure time-windows: the whole pregnancy, and each trimester of the pregnancy (92 days per trimester if no delivery). We calculated the exposure of women who moved during a time-window (11.8%) using time-weighted averages of levels at each home address. 2.2.1. Nearest AQMS model We obtained NO2 and PM10 hourly concentrations from the permanent AQMS (Fig. 1a, e, i, k) and considered circular buffers with a 10 km radius around each AQMS for PM10 and 5 km for NO2 (see Supplemental Material). We also studied buffers of 2 km and 1 km. AQMSs located up to within 2.5 km of the limits of the study area were included, except traffic (i.e., located b5 m from a road with traffic levels of N10,000 vehicles/day) (ADEME, 2002) and industrial AQMS. 2.2.2. Temporally-adjusted geostatistical model NO2 measurement campaigns were conducted in Nancy in 2002 and Poitiers in 2005 and repeated throughout the year so as to capture seasonal variations (see Supplemental Material). The campaigns were conducted in the metropolitan areas of both cities using Passam tubes and Palmes diffusive samplers (Palmes et al., 1976). The tubes were located so as to give measurements of air pollution levels in urban, suburban and rural areas (98 locations in Nancy, 61 locations in Poitiers). The estimated annual concentrations were smoothed over the whole area using universal kriging on a 50 × 50 m grid using Isatis Software version 6.06 (Geovariances, Fontainebleau, France) (Fig. 1b, f). These spatiallyresolved estimates were then combined to time-specific measurements from AQMSs to capture temporal variation relevant to each timewindow (Lepeule et al., 2010). 2.2.3. Land-use regression model From the annual NO2 averages estimated from the measurement campaigns described above, we built a LUR model following the methodology of the ESCAPE (European Study of Cohorts for Air Pollution Effects) project and reviewed by its exposure working group (Fig. 1d, h) (Beelen R et al., unpublished data) (see Supplemental Material). From the 98 locations measured in Nancy and 61 measured in Poitiers, respectively, 39 and 7 locations were discarded to meet the requirements of the ESCAPE protocol, and 1 and 3 locations were discarded because they were influential observations (according to Cook's D value), leaving 58 locations in Nancy and 51 in Poitiers. The LUR equations included the following concentration predictors: total traffic load of major roads (1 km), length of major roads (25 m), and length of all roads (50 m) in Nancy; total traffic load of all roads (within 25 m), number of inhabitants (300 m), presence of industry (300 m), and length of major roads (1 km) in Poitiers. Cross-validation R2 was 0.67 in Nancy and 0.80 in Poitiers. These equations were applied to each woman's address. As for the TAG model, these spatially-resolved estimates were combined to AQMS time-specific measurements.
167
2.2.4. Dispersion model NO2 and PM10 hourly concentrations were modeled at women's home addresses using the quasi-Gaussian Atmospheric Dispersion Modeling System (ADMS) Urban 3.1 (Cambridge Environmental Research Consultants, Cambridge, United Kingdom) (AIRLOR, 2010; Galineau et al., 2011) (see Supplemental Material). Pearson correlations between the hourly model predictions and the measures at the AQMS locations ranged from 0.50 to 0.65 for NO2 in Nancy and 0.58 to 0.62 for PM10. In Poitiers, the correlations ranged from 0.65 to 0.70 for NO2 and 0.85 to 0.95 for PM10. Model outputs for annual mean concentrations in NO2 and PM10 in 2005 were predicted (Isatis Software, Géovariances, Fontainebleau) using a regular grid with points every 300 m and the ADMS intelligent gridding setting, which further places 2 points on either side of the road distributed every 100 m along the road. Annual concentrations were mapped on 50 × 50 m interpolation grid (Fig. 1c, g, j, l). 2.2.5. Study area For each pollutant, the study area was the geographic area at the intersection between all exposure models (Fig. 1). The study area included urban (city), suburban (metropolitan area according to the regional authorities' classification of communities) and rural areas. 2.3. Statistical analysis We performed variance decomposition of the PM10 nearest AQMS and NO2 LUR models (not possible for the dispersion model) to assess their spatial and temporal components (Lepeule et al., 2010). For each pair of exposure models, we compared exposure estimates using the Wilcoxon signed rank test, Spearman rank correlation (r), and Bland– Altman plots adapted for more than two measures when there is no gold standard (Seed, 2000). We calculated Spearman correlation by city, type of area, buffer size and exposure time-window. The relationship between exposure estimated by each exposure model and birth weight was estimated by separate linear regressions adjusted for the following a priori-chosen covariates: gestational age (linear and quadratic), sex of the newborn, maternal height, pre-pregnancy weight (broken stick regression with a knot at 60 kg), birth order, maternal age at the end of education, city, month of conception, and active and passive smoking during the relevant time-window under study. We examined effect modification by city by including an interaction term with the exposure variable. The exposure–response relationships were further assessed for each buffer size. Time-windows of exposure were first examined separately. We then simultaneously adjusted for the three trimesters of pregnancy. Results were considered statistically significant when p-value b 0.05. Statistical analyses were carried out with SAS version 9.2 (SAS Institute, Cary, NC). 3. Results 3.1. Study area and study sample The PM10 study area included 1026 women and the NO2 study area included 776 women. Women mainly lived in urban and suburban settings (Table 1). For most characteristics, mothers living within different distances from a monitor did not clearly differ (see Supplemental Material, Tables S1 and S2). 3.2. Comparison of exposure estimates according to the exposure model As results by trimester of the pregnancy for both pollutants were similar to those for the whole pregnancy, we only describe the latter. Depending on the exposure model, mean pregnancy concentrations were between 22 and 29 μg/m3 for NO2 and between 20 and 21 μg/m3 for PM10. All exposure models showed lower exposure levels in rural than in urban and suburban areas (Fig. 2) and lower averages in Poitiers
168
NO2
b) POITIERS : Geostatistical model
e) NANCY : Nearest AQMS model
f) NANCY : Geostatistical model
c) POITIERS : Dispersion model
d) POITIERS : Land-use regression
g) NANCY : Dispersion model
h) NANCY : Land-use regression
i) POITIERS : Nearest AQMS model
j) POITIERS : Dispersion model
PM10
k) NANCY : Nearest AQMS model
l) NANCY : Dispersion model
Fig. 1. NO2 (a–h) and PM10 (i–l) concentration levels in the metropolitan area of Poitiers and Nancy in 2005 estimated by four models: a, e, i, k) nearest AQMS (air quality monitoring station) model; b, f) geostatistical model; c, g, j, l) dispersion model; and d, h) land-use regression model.
Y. Sellier et al. / Environment International 66 (2014) 165–173
a) POITIERS : Nearest AQMS model
Y. Sellier et al. / Environment International 66 (2014) 165–173
169
32.3±10 24.9±5 28.1±12 23.0±8
25.5±9 23.0±4 25.8±10 21.1±6
16.7±3 17.9±3 16.3±4 12.3±4
20.8±3
20.7±3
20.3±4
21.1±4
20.4±5
20.5±3
20.8±4
19.1±4
35
28.6±10 23.7±5 26.5±11 21.7±7
0
0
5
10
10
20
15
30
20
25
40
30
50
60
AQMS TAG LUR Dispersion
776 770 776 776
Whole area
414 412 414 414
320 319 320 320
Urban area
Suburban area
42
42
39
42
1025
Rural area
1026
Whole area
424
424
Urban area
NO2
377
377
224
Suburban area
n
225
Rural area
PM10
area
Pollutant
Fig. 2. Box plots and average ± SD of maternal pregnancy exposure to NO2 and PM10 (μg/m3) estimated by four exposure models, EDEN mother–child cohort. Boxes represent the 25th, 50th, and 75th percentiles. Whiskers represent the lower quartile–1.5 interquartile range and the upper quartile + 1.5 interquartile range.
compared to Nancy (see Supplemental Material, Fig. S1). For both pollutants, all pairwise comparisons of concentration distributions differed between exposure models (p-values b 0.001).
correlation among exposure models other than the nearest AQMS was not affected by the buffer size. When both cities were considered separately, the correlations with the nearest AQMS were much lower in Poitiers compared to Nancy (see Supplemental Material, Table S3). Compared to other models, the nearest AQMS tended to underestimate NO2 exposure when the concentration was low and to overestimate when the concentration was high (see Supplemental Material, Fig. S2). To a lesser extent, we observed the opposite trend for the TAG model. Since the LUR was the only model to capture exposure estimates N45 μg/m3 (Fig. 2), the scatter plot indicated an overestimation from this model for high concentrations (see Supplemental Material, Fig. S2). The dispersion model seemed generally to underestimate the exposure levels.
3.2.1. Nitrogen dioxide Pregnancy exposure estimates were subject to stronger spatial variations for the seasonalized LUR (accounting for 99% of the variance of exposure) than for the nearest AQMS (95%) and TAG (92%) models. The spatial contribution was lower when we considered trimesterspecific windows with 76–87% for the LUR, 78–82% for the nearest AQMS and 47–61% for the TAG. As for the whole area, the range of exposure estimates was largest for the LUR and dispersion models, and lowest for the TAG model (Fig. 2). Mean exposure decreased from the nearest AQMS, then the LUR, the TAG and finally to the dispersion models (for the whole area of the study). Correlations between model exposure estimates were high (≥0.70), except for the nearest AQMS with both the LUR (r = 0.54) and dispersion models (r = 0.63, Table 2). The highest correlation was between the LUR and dispersion with r = 0.87. Nearest AQMS better correlated with other models when we restricted the population to women living within 2 km to 1 km from the nearest AQMS (Table 2). As expected, the
3.2.2. Particles The spatial component of the nearest AQMS accounted for 22% of the variation of PM 10 exposure during pregnancy and for 15– 19% by trimester of the pregnancy. The mean exposure level produced by the nearest AQMS and dispersion models was similar, although the dispersion model exhibited a larger range (Fig. 2). The correlation between the nearest AQMS and the dispersion model
Table 2 Correlation between pairs of four exposure models estimating NO2 and PM10 levels during pregnancy, EDEN mother–child cohort. Pollutant
Exposure models
10 n NO2
AQMS-
TAG-
PM10
LURAQMS-
Type of areaa
Buffer size of the nearest AQMS (km)
TAG LUR Dispersion LUR Dispersion Dispersion Dispersion
1025
5
2
1
Urban
Suburban
Rural
r
n
r
n
r
n
r
n
r
n
r
n
r
0.81
770 776 776 770 770 776 790
0.70 0.54 0.63 0.79 0.79 0.87 0.81
428 431 431 428 428 431 372
0.76 0.59 0.65 0.78 0.75 0.86 0.85
157 158 158 157 157 158 136
0.85 0.76 0.71 0.77 0.63 0.84 0.85
412 414 414 412 412 414 424
0.73 0.54 0.62 0.78 0.75 0.86 0.86
319 320 320 319 319 320 377
0.51 0.46 0.53 0.73 0.75 0.84 0.75
39 42 42 39 39 42 224
0.66 0.59 0.69 0.87 0.73 0.77 0.83
AQMS: air-quality monitoring station model (home's nearest AQMS in a 5 km buffer for NO2 and 10 km buffer for PM10), TAG: temporally adjusted geostatistical model, LUR: land use regression model, n: sample size, r: Spearman correlation. a Using a 5 km buffer around each AQMS for NO2 and a 10 km buffer for PM10.
170
Y. Sellier et al. / Environment International 66 (2014) 165–173
each 10 μg/m3 increase in PM10 exposure as estimated by the dispersion model during the 1st trimester of pregnancy, and a 175 g (95%CI: 48; 301) increase in birth weight associated with each 10 μg/m3 increase in PM10 exposure during the 2nd trimester of pregnancy.
was 0.81, a value higher than for NO2. The correlation improved a little between 5 km and 2 km (Table 2) and was higher in Poitiers (r = 0.73) than in Nancy (r = 0.44) (see Supplemental Material, Table S3). The nearest AQMS tended to provide higher PM10 concentration both at low and high levels (see Supplemental Material, Fig. S2).
4. Discussion 3.3. Effect estimates of NO2 and PM10 on birth weight Fetal life is a sensitive period during which exposure to toxicants may have immediate and also delayed adverse effects on health (Perera and Herbstman, 2011). Because timing of exposure is likely to be a key feature for explaining the health effects of fetal exposure to toxicants, both spatial and temporal exposure variability require to be captured by exposure metrics. We developed four exposure models with different spatial and temporal resolution, and compared their exposure estimates for pregnant women and effect estimates on infant birth weight. For PM10, the nearest AQMS and dispersion models indicated consistent results both in terms of exposure estimates and association with birth weight. For NO2, although there was an overall moderate to good correlation among the four exposure models (r in the 0.54–0.87 range), the direction of the association with birth weight differed according to the model used for estimating air pollution exposure.
NO2 effect estimates on birth weight generally had large confidence intervals centered around zero and the sign of the point estimate varied according to the exposure model (Fig. 3). As for the exposure timewindow, considering results from the 4 exposure models, there was no clear pattern of sensitivity to NO2 through the pregnancy. PM10 levels were positively associated with birth weight, with significant associations during the second trimester of pregnancy as estimated by the dispersion model. For both pollutants, we did not find any effect modification by city. As for uncertainty, for both pollutants, health effect estimates based on the dispersion and TAG models had the largest variance, while those from the AQMS and LUR models provided the lowest variance. When restricting to smaller buffer size around AQMS, the effect estimates of NO2 on birth weight decreased for all exposure models and all time-windows. We observed the opposite trend for PM10. Further adjusting for all exposure time-windows (correlations ≤ 0.86, see Supplemental Material, Table S4) changed subsequently the health effect estimates (see Supplemental Material, Fig. S3). For all exposure models and for both pollutants, the association with birth weight tended to be negative with exposure during the 1st trimester of pregnancy, positive with the 2nd trimester of pregnancy and null with the 3rd trimester of pregnancy. These associations were stronger for PM10 than for NO2, with a 90 g (95%CI: −220; 42) decrease in birth weight associated with
4.1. Characteristics and agreement of exposure models For NO2, we observed good correlations (i.e. ≥ 0.79) between the TAG, LUR and dispersion models, while the nearest AQMS exhibited moderate correlations from 0.54 to 0.70 with the aforementioned models. Geostatistical models are generally more spatially resolved than AQMS-based models, but do not capture fine within-urban variations in air pollution due to local sources. Therefore, geostatistical and
NO2 AQMS TAG LUR DISPERSION
Change in birth weight in gr
200 150 100 50 0 -50 -100 -150 -200 -250
5km
2km
1km
5km
2km
1km
5km
2km
1km
5km
2km
(704)
(414)
(164)
(703)
(405)
(154)
(704)
(398)
(147)
(706)
(395)
1st trimester
2nd trimester
3rd trimester
Pregnancy
1km Buffer size (147) (n) Exposure window
PM10 Change in birth weight in gr
500
AQMS
DISPERSION
400 300 200 100 0 -100 -200 -300 -400
10km
5km
2km
1km
10km
5km
2km
1km
10km
5km
2km
1km
10km
5km
2km
(926)
(730)
(364)
(147)
(890)
(688)
(337)
(132)
(891)
(692)
(332)
(126)
(940)
(724)
(341)
1st trimester
2nd trimester
3rd trimester
Pregnancy
1km Buffer size (126) (n) Exposure window
Fig. 3. Change in mean birth weight (g) for a 10 μg/m3 increase in air pollutant concentration, as a function of the size of the buffer considered around each AQMS, adjusted for factors as described in “Methods” section. Error bars indicate 95%CIs.
Y. Sellier et al. / Environment International 66 (2014) 165–173
AQMS models are usually thought to represent regional air quality. Accordingly, exposure estimates from our TAG model exhibited the lowest variability (range and SD) of the four exposure models. LURs allow capturing within-urban variability in air pollutant concentrations, which is illustrated by the stronger contribution of the spatial component in the LUR compared to the TAG model. Dispersion models simulate the chemical reactions, transport and dispersion of pollutants and can capture fine contrasts in both spatial and temporal scales thanks to appropriate input data. In our study, the TAG model was well correlated with both the LUR and dispersion models (r = 0.79), which confirms its ability to capture some of the spatial variability in NO2 concentrations. The LUR and dispersion models, which were highly correlated (r = 0.86), provided highly spatially-resolved estimates (Fig. 1). 4.2. Influence of the location of AQMS and sources of air pollution Exposure estimates from AQMS-based models strongly depend on the location and density of monitors, as previously extensively discussed (e.g., Lepeule et al., 2010). The somewhat different ranges and correlations of NO2 levels between the nearest AQMS and other models in Poitiers (see Supplemental Material, Table S3, Fig. S1) are likely due to the influence of the city-center AQMS, which experienced very high NO2 concentrations (Lepeule et al., 2010, Fig. 2, station M). More generally, the stronger improvement of the correlation between the nearest AQMS and the other models for NO2 from 2 to 1 km (6– 16 units, Table 2) compared to 5 to 2 km (2–6 units) also reflects the fact that AQMS is usually only representative of air pollution levels in a very small surrounding area (Lebret et al., 2000). As for PM10, the correlation between the nearest AQMS and dispersion models was not substantially influenced by the buffer size around the monitors. These results likely own to the quite homogenous distribution of PM10 concentrations across the study area, which is illustrated by the low spatial variability of the nearest AQMS model and is in line with the results obtained in European countries (Gulliver et al., 2011). When the buffer size of the AQMS model decreased, effect estimates moved usually further from the null (except for the PM10 1 km buffer). These results are unlikely to be due to a selection bias (see Supplementary Materials, Tables S1 and S2) but rather result from a decrease in home address exposure misclassification. PM10 in Nancy and Poitiers originate from both regional pollution episodes and local sources such as traffic and residential emissions. Pollutant concentrations were higher in Nancy, which is more urbanized and densely populated compared to Poitiers. In both cities, NO2 concentrations were higher in urban and suburban areas, where traffic, urban heating and industries are the main sources, compared to rural areas.
171
ours. In our study, the LUR model predicted the highest concentrations, especially in Nancy where exposures between 45 and 73 μg/m3 were estimated during pregnancy. This result is in agreement with those of Marshall et al. (2008) and the levels estimated in Nancy are compatible with its urbanization and the measures of monitors performed in traffic proximity. In the Vancouver study, the authors reported significant positive associations between NO2 exposure during pregnancy and term low birth weight (b2500 g) when they used an AQMS-based approach, but no association when they used the LUR (Brauer et al., 2008). Although we did not find any significant association of continuous birth weight with whole pregnancy exposure, the pattern of our results (stronger effects with the nearest AQMS than with the LUR) is compatible with this study. Wu et al. (2011) estimated NO2 during pregnancy and found a correlation of 0.57 between a temporally-adjusted LUR and the nearest AQMS. They reported a positive association of preterm birth with the nearest AQMS exposure, but a negative association with the LUR estimate. In that same study, Laurent et al. reported positive associations between all pollutants (except ozone) measured by the nearest AQMS, a LUR or a dispersion model and term birth weight (Laurent et al., 2013). In a study estimating air pollution exposure at home and workplace during pregnancy, Madsen et al. (2010) reported a very low correlation between exposures as estimated by a fine spatial scale dispersion model and the fixed monitor of the city (0.07 for NO2 and 0.19 for PM10) and found negative associations with birth weight for both pollutants as estimated by the monitor, but positive associations using the dispersion model. Although we do not have any biological plausible explanation for these positive associations with birth weight, our results agree with those studies. However, such results are not corroborated by a recent meta-analysis (Dadvand et al., 2013) indicating adverse effects of whole pregnancy exposure to particles on term birth weight. In comparison, the lack of significance of our results regarding whole pregnancy NO2 and birth weight may be due to our limited statistical power. It might also be viewed as a consistent result among exposure models indicating no significant association with continuous birth weight, perhaps because NO2 is not a relevant marker for air pollution effects on birth weight. In our sensitivity analyses adjusted for all trimesters of exposure, we saw, for all exposure models and both pollutants, a trend for a negative association of birth weight with exposure during the first trimester of pregnancy, while there was a trend for a positive association with the second trimester of pregnancy. Such varying effects of exposure during the different trimesters of pregnancy might be artifacts due to the use of attained weight at birth, which does not represent the patterns of fetal growth along the three trimesters. A way to overcome this issue would be to use ultrasound measurements during the course of the pregnancy (Aguilera et al., 2010; van den Hooven et al., 2012).
4.3. Comparison with the epidemiological literature 4.4. Exposure assessment and time–activity patterns Few papers have compared different methods for estimating air pollution exposure in terms of impact on health effect estimates (Laurent et al., 2013; Sahsuvaroglu et al., 2009; Son et al., 2010). As reproductive epidemiology often deals with exposure windows of 9 months or less, both spatial and temporal resolution exposure contrasts should be considered. Our results regarding the nearest AQMS and TAG NO2 models have been previously discussed (Lepeule et al., 2010). We therefore focus here on comparisons involving the fine spatial scale LUR and dispersion models as well as PM10. In a term-born infant study in California, Laurent et al. found NO2 correlations of 0.57 between the nearest AQMS and LUR (Laurent et al., 2013). In a comparison study of annual NO2 levels in Vancouver, Marshall et al. (2008) found correlations of 0.54 between the nearest AQMS and LUR, 0.61 between the nearest AQMS and dispersion models and 0.49 between the LUR and dispersion models, which are comparable to our results (0.54, 0.63, and 0.87 respectively). They used a larger 10 km buffer around monitors and their dispersion model accounted for downtown versus suburban variations only, which likely explains its low correlation with the LUR compared to
Another explanation for the inconsistent relationships with birth weight among the four exposure models and for the unexpected positive relationships lies in the exposure measurement error. As in most epidemiological studies, we considered air pollution exposure of women at their home address as a surrogate of the personal exposure, which includes indoor and outdoor sources and transportations, as a function of time spent in these microenvironments. In the Vancouver study, Nethery et al. (2008) evaluated LUR estimates as predictors of personal exposure during pregnancy. They reported that accounting for work outdoor concentrations improved LUR estimates, whose correlation with personal exposure was 0.28, compared to 0.05 when the LUR model focused on home address only. However, these results suggest that there is ample room for improvements of air pollution exposure estimates. For example, time spent commuting was shown to be responsible for 29% of personal benzene exposure (Bruinen de Bruin et al., 2008). Highly spatially resolved air pollution models such as our LUR and dispersion models predict levels that are very specific to the
172
Y. Sellier et al. / Environment International 66 (2014) 165–173
location, usually the subject's home address, while exposure models relying on monitors or interpolation methods provide a mean level over a broader area, which may include other microenvironments of the subject. Therefore, not accounting for time–activity patterns when using highly spatially-resolved exposure models may create a similar or bigger exposure misclassification than, for example, the error made by the nearest AQMS model, and therefore produce biased associations with birth weight.
FRM, Inserm, IReSP, Nestlé, French Ministry of Health, ANR, Univ. Paris-Sud, InVS, ANSES and MGEN. The Team of Environmental Epidemiology at Inserm U823 is supported by an AVENIR/ATIP grant. The authors declare that they have no competing financial interests. Appendix A. Supplementary data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.envint.2014.02.001.
4.5. Generalizability of the results and relevance of exposure models References A recent study comparing four PM10 exposure models in London indicated poor performance of the nearest AQMS and kriging (both based on 52 monitors) to predict long-term ambient PM and slightly better performances of the LUR and dispersion models (Gulliver et al., 2011). The authors concluded that these models rather predict the average level of the urban area than the local patterns of variation and that LUR might be the most robust method to predict long-term exposure in complex areas. However, the generalizability of results from exposure model comparison is limited to areas with similar characteristics, such as regional and local sources of emissions, monitor coverage, and weather and topography of the area. Furthermore, the relevance of the method used to assess exposure depends on the spatio-temporal variability of the pollutants, the time-window length and endpoints considered and the mobility of the population. For example, assessment of exposure to PM10, due to its usually low spatial variability, may not be much improved by the use of highly-spatially resolved methods such as our dispersion model, which is computationally intensive and requires data that may not be easily available everywhere. The LUR is thought to perform at least as well as a dispersion model when a sufficient dense network of monitors is used (Dijkema et al., 2011). In ESCAPE, LUR models relying on 40 NO2 measurement sites have been found efficient as long as the measurement sites included a large range of potential sources of air pollution variability (Beelen R et al., unpublished data). Other studies consider that a larger number is required (Basagana et al., 2012). Our LUR included at least 51 sites in each city distributed so as to represent such variability. However, one should bear in mind that LUR is not initially designed for reflecting temporal variations and may not account for the meteorology, which might be an issue for certain areas. Fixed monitor data are easily accessible, usually free and provide high temporally resolved models with however a usually poor spatial resolution, which has led to the conclusion that these exposure models are likely to introduce exposure error and bias the effect estimates. However, studies such as short-term health effects might still benefit from monitoring data in certain conditions (Sarnat et al., 2010). 5. Conclusion Our study adds to the growing body of literature on maternal air pollution exposure effects on fetal growth by showing that different spatiotemporal approaches of air pollution exposure assessment, although with well correlated predictions, provide different rankings of the subjects regarding their exposure and may lead to different conclusions regarding associations with health endpoints. Accounting for time– activity patterns of the subjects rather than solely considering the home addresses might be a key issue for highly-spatially resolved exposure models compared to models based on between-community comparisons. Acknowledgments We are indebted to the midwife research assistants (L. Douhaud, S. Bedel, B. Lortholary, S. Gabriel, M. Rogeon, and M. Malinbaum) for data collection and to P. Lavoine for checking, coding, and data entry. This work was supported by ANSES. The EDEN cohort is funded by
ADEME. Classification and criteria for setting up air-quality monitoring stations. Paris ADEME ÉditionsParis ADEME Éditions; 2002. Aguilera I, Garcia-Esteban R, Iniguez C, Nieuwenhuijsen MJ, Rodriguez A, Paez M, et al. Prenatal exposure to traffic-related air pollution and ultrasound measures of fetal growth in the INMA Sabadell cohort. Environ Health Perspect 2010;118:705–11. AIRLOR. Concentration en polluants dans l'air ambiant extérieur aux domiciles des sujets de la cohorte épidémiologique EDEN. AIRLORNancy: AIRLOR; 2010. Basagana X, Rivera M, Aguilera I, Agis D, Bouso L, Elosua R, et al. Effect of the number of measurement sites on land use regression models in estimating local air pollution. Atmos Environ 2012;54:634–42. Brauer M, Lencar C, Tamburic L, Koehoorn M, Demers P, Karr C. A cohort study of traffic-related air pollution impacts on birth outcomes. Environ Health Perspect 2008;116:680–6. Briggs D. The role of GIS: coping with space (and time) in air pollution exposure assessment. J Toxicol Environ Health A 2005;68:1243–61. Bruinen de Bruin Y, Koistinen K, Kephalopoulos S, Geiss O, Tirendi S, Kotzias D. Characterisation of urban inhalation exposures to benzene, formaldehyde and acetaldehyde in the European Union: comparison of measured and modelled exposure data. Environ Sci Pollut Res Int 2008;15:417–30. Cyrys J, Hochadel M, Gehring U, Hoek G, Diegmann V, Brunekreef B, et al. GIS-based estimation of exposure to particulate matter and NO2 in an urban area: stochastic versus dispersion modeling. Environ Health Perspect 2005;113:987–92. Dadvand P, Parker J, Bell ML, Bonzini M, Brauer M, Darrow LA, et al. Maternal exposure to particulate air pollution and term birth weight: a multi-country evaluation of effect and heterogeneity. Environ Health Perspect 2013;121:267–373. Dijkema MB, Gehring U, van Strien RT, van der Zee SC, Fischer P, Hoek G, et al. A comparison of different approaches to estimate small-scale spatial variation in outdoor NO(2) concentrations. Environ Health Perspect 2011;119:670–5. Galineau J, Hulin A, Caini F, Marquis N, Lepeule J, Slama R. Estimation of exposure to urban air pollution in two cities using a Gaussian dispersion model: the EDEN-Air project. Abstracts of the 23rd Annual Conference of the International Society of Environmental Epidemiology (ISEE) September 13–16, 2011, Barcelona, SpainEnviron Health Perspect; 2011. Gulliver J, de Hoogh K, Fecht D, Vienneau D, Briggs D. Comparative assessment of GIS-based methods and metrics for estimating long-term exposures to air pollution. Atmos Environ 2011;45:7072–80. Jerrett M, Burnett RT, Ma R, Pope III CA, Krewski D, Newbold KB, et al. Spatial analysis of air pollution and mortality in Los Angeles. Epidemiology 2005;16:727–36. Laurent O, Wu J, Li L, Chung J, Bartell S. Investigating the association between birth weight and complementary air pollution metrics: a cohort study. Environ Health 2013;12:18. Lebret E, Briggs D, van Reeuwijk H, Fischer P, Smallbone K, Harssema H, et al. Small area variations in ambient NO2 concentrations in four European areas. Atmos Environ 2000;34:177–85. Lepeule J, Caini F, Bottagisi S, Galineau J, Hulin A, Marquis N, et al. Maternal exposure to nitrogen dioxide during pregnancy and offspring birth weight: comparison of two exposure models. Environ Health Perspect 2010;118:1483–9. Madsen C, Gehring U, Walker SE, Brunekreef B, Stigum H, Naess O, et al. Ambient air pollution exposure, residential mobility and term birth weight in Oslo, Norway. Environ Res 2010;110:363–71. Marshall JD, Nethery E, Brauer M. Within-urban variability in ambient air pollution: comparison of estimation methods. Atmos Environ 2008;42:1359–69. Nethery E, Leckie SE, Teschke K, Brauer M. From measures to models: an evaluation of air pollution exposure assessment for epidemiological studies of pregnant women. Occup Environ Med 2008;65:579–86. Palmes ED, Gunnison AF, DiMattio J, Tomczyk C. Personal sampler for nitrogen dioxide. Am Ind Hyg Assoc J 1976;37:570–7. Pereira G, Nassar N, Cook A, Bower C. Traffic emissions are associated with reduced fetal growth in areas of Perth, Western Australia: an application of the AusRoads dispersion model. Aust N Z J Public Health 2011;35:451–8. Perera F, Herbstman J. Prenatal environmental exposures, epigenetics, and disease. Reprod Toxicol 2011;31:363–73. Pouliou T, Kanaroglou PS, Elliott SJ, Pengelly LD. Assessing the health impacts of air pollution: a re-analysis of the Hamilton children's cohort data using a spatial analytic approach. Int J Environ Health Res 2008;18:17–35. Rahmalia A, Giorgis-Allemand L, Lepeule J, Philippat C, Galineau J, Hulin A, et al. Pregnancy exposure to atmospheric pollutants and placental weight: an approach relying on a dispersion model. Environ Int 2012;48:47–55. Rosenlund M, Forastiere F, Stafoggia M, Porta D, Perucci M, Ranzi A, et al. Comparison of regression models with land-use and emissions data to predict the spatial distribution of traffic-related air pollution in Rome. J Expo Sci Environ Epidemiol 2008;18: 192–9.
Y. Sellier et al. / Environment International 66 (2014) 165–173 Sahsuvaroglu T, Jerrett M, Sears MR, McConnell R, Finkelstein N, Arain A, et al. Spatial analysis of air pollution and childhood asthma in Hamilton, Canada: comparing exposure methods in sensitive subgroups. Environ Health 2009;8:14. Sarnat SE, Klein M, Sarnat JA, Flanders WD, Waller LA, Mulholland JA, et al. An examination of exposure measurement error from air pollutant spatial variability in time-series studies. J Expo Sci Environ Epidemiol 2010;20:135–46. Seed P. sbe33: comparing several methods of measuring the same quantity. Stata Tech Bull 2000;55:2–9. Slama R, Morgenstern V, Cyrys J, Zutavern A, Herbarth O, Wichmann HE, et al. Traffic-related atmospheric pollutants levels during pregnancy and offspring's term birth weight: a study relying on a land-use regression exposure model. Environ Health Perspect 2007;115:1283–92.
173
Slama R, Thiebaugeorges O, Goua V, Aussel L, Sacco P, Bohet A, et al. Maternal personal exposure to airborne benzene and intrauterine growth. Environ Health Perspect 2009;117:1313–21. Son JY, Bell ML, Lee JT. Individual exposure to air pollution and lung function in Korea: spatial analysis using multiple exposure approaches. Environ Res 2010;110:739–49. van den Hooven EH, Pierik FH, de Kluizenaar Y, Willemsen SP, Hofman A, van Ratingen SW, et al. Air pollution exposure during pregnancy, ultrasound measures of fetal growth, and adverse birth outcomes: a prospective cohort study. Environ Health Perspect 2012;120:150–6. Wu J, Wilhelm M, Chung J, Ritz B. Comparing exposure assessment methods for traffic-related air pollution in an adverse pregnancy outcome study. Environ Res 2011;111:685–92.
Contents lists available at ScienceDirect
Environment International journal homepage: www.elsevier.com/locate/envint
Health effects of ambient air pollution: Do different methods for estimating exposure lead to different results? Yann Sellier a,b, Julien Galineau c, Agnes Hulin d, Fabrice Caini d, Nathalie Marquis c, Vladislav Navel d, Sebastien Bottagisi a,b, Lise Giorgis-Allemand a,b, Claire Jacquier c, Remy Slama a,b, Johanna Lepeule a,b,e,⁎, the EDEN mother–child cohort study group a
Inserm, U823, Institut Albert Bonniot, Team of Environmental Epidemiology Applied to Reproduction and Respiratory Health, Grenoble, France Université Joseph Fourier, Grenoble, France c Air Lorraine, Nancy, France d ATMO Poitou-Charentes, La Rochelle, France e Exposure, Epidemiology, and Risk Program, Department of Environmental Health, Harvard School of Public Health, Boston, MA 02115, USA b
a r t i c l e
i n f o
Article history: Received 27 June 2013 Accepted 4 February 2014 Available online 2 March 2014 Keywords: Exposure modeling Measurement error Spatial resolution
a b s t r a c t Background: Spatially resolved exposure models are increasingly used in epidemiology. We previously reported that, although exhibiting a moderate correlation, pregnancy nitrogen dioxide (NO2) levels estimated by the nearest air quality monitoring station (AQMS) model and a geostatistical model, showed similar associations with infant birth weight. Objectives: We extended this study by comparing a total of four exposure models, including two highly spatially resolved models: a land-use regression (LUR) model and a dispersion model. Comparisons were made in terms of predicted NO2 and particle (aerodynamic diameter b 10 μm, PM10) exposure and adjusted association with birth weight. Methods: The four exposure models were implemented in two French metropolitan areas where 1026 pregnant women were followed as part of the EDEN mother–child cohort. Results: Correlations between model predictions were high (≥0.70), except for NO2 between the AQMS and both the LUR (r = 0.54) and dispersion models (r = 0.63). Spatial variations as estimated by the AQMS model were greater for NO2 (95%) than for PM10 (22%). The direction of effect estimates of NO2 on birth weight varied according to the exposure model, while PM10 effect estimates were more consistent across exposure models. Conclusions: For PM10, highly spatially resolved exposure model agreed with the poor spatial resolution AQMS model in terms of estimated pollutant levels and health effects. For more spatially heterogeneous pollutants like NO2, although predicted levels from spatially resolved models (all but AQMS) agreed with each other, our results suggest that some may disagree with each other as well as with the AQMS regarding the direction of the estimated health effects. © 2014 Elsevier Ltd. All rights reserved.
1. Introduction Great advances have been achieved over the past 5 years in improving the spatial resolution of air pollution exposure models used for studying its mid- and long-term health effects. Epidemiological studies have moved from an approach based on air-quality monitors, which provides temporally resolved estimates but is unlikely to capture the within-city spatial heterogeneity of air pollutant concentrations (Lebret et al., 2000), to more spatially resolved models based either on measured concentrations, combined with geographical information ⁎ Corresponding author at: Equipe 12, Centre de Recherche INSERM-UJF U823, Institut Albert Bonniot, UJF Site Santé, BP 170 La Tronche, 38042 GRENOBLE Cedex 9, France. Tel.: +33 476 549 466. E-mail address: [email protected] (J. Lepeule).
http://dx.doi.org/10.1016/j.envint.2014.02.001 0160-4120/© 2014 Elsevier Ltd. All rights reserved.
system data (land-use regression) and/or geostatistical techniques, or on dispersion modeling (Briggs, 2005). A few studies have compared the performance of such exposure models and concluded to the existence of substantial differences (Cyrys et al., 2005; Gulliver et al., 2011; Marshall et al., 2008; Rosenlund et al., 2008). For endpoints such as mortality and respiratory health, measurement error resulting from the use of poorly spatially resolved models may have a large impact on the exposure–response relationship (Jerrett et al., 2005; Pouliou et al., 2008). Although fine spatial scale land-use regression and dispersion models are increasingly used in reproductive epidemiology (Aguilera et al., 2010; Pereira et al., 2011; Rahmalia et al., 2012; Slama et al., 2007), few systematic comparisons have been undertaken in terms of exposure estimates (Marshall et al., 2008) or health effect estimates (Brauer et al., 2008; Lepeule et al., 2010; Wu et al., 2011). Since implementing such fine scale exposure models is costly, and since air
166
Y. Sellier et al. / Environment International 66 (2014) 165–173
quality monitoring data are widely available, there is a critical need for understanding the implications of the use of such exposure models. The well-characterized EDEN mother–child cohort provides unique opportunities to study up to four exposure models to nitrogen dioxide (NO2) and particles with aerodynamic diameter less than 10 μm (PM10) in two metropolitan areas: the nearest air quality monitoring station (AQMS) model, a temporally-adjusted geostatistical (TAG) model, a land-use regression (LUR) model and a dispersion model. We previously showed that NO2 levels during pregnancy estimated by the nearest AQMS and by the TAG were moderately correlated; nevertheless they exhibited similar associations with infant birth weight (Lepeule et al., 2010). In the present study, we extended our work by considering the LUR and dispersion models, aiming at capturing fine
spatial contrasts in air pollutant concentrations and by considering PM10 estimated by the nearest AQMS and dispersion models. We compared these four NO2 exposure models and two PM10 exposure models in terms of exposure estimates during pregnancy and we investigated how the relationship with birth weight varied according to the exposure model used. 2. Methods 2.1. Study population and data collection The EDEN mother–child cohort consists in 2002 women enrolled before 26 gestational weeks at maternity wards of Nancy and Poitiers
Table 1 Characteristics of the study population, EDEN mother–child cohort. Characteristic
NO2 study area (n = 776)
PM10 study area (n = 1026)
Birth weight, g, mean (sd) Sex of the offspring, n (%) Male Female Gestational duration (weeks), n (%) 30–36 37–38 39–40 ≥41 Birth order, n (%) First Second Third or more Missing Month of conception of the child, n (%) January–March April–June July–September October–December Maternal age at conception (years), n (%) b25 25–29 30–34 ≥35 Maternal height (cm), n (%) b160 160–169 ≥170 Missing Maternal pre-pregnancy weight (kg), n (%) b50 50–59 60–69 70–79 ≥80 Missing Center, n (%) Poitiers Nancy Urbanization, n (%) Urban Suburban Rural Maternal age at end of education (years), n (%) ≤16 17–18 19–20 21–22 23–24 ≥25 Maternal active smoking (2nd trimester), n (%) No Yes Missing Maternal passive smoking (2nd trimester), n (%) No Yes Missing
3284 (512)
3278 (505)
395 (50.9) 381 (49.1)
531 (51.7) 495 (48.3)
47 (6.1) 128 (16.5) 440 (56.7) 161 (20.7)
60 (5.9) 176 (17.1) 580 (56.5) 210 (20.5)
367 (47.3) 263 (33.9) 145 (18.7) 1 (0.1)
477 (46.5) 365 (35.6) 182 (17.7) 2 (0.2)
167 (21.5) 184 (23.7) 226 (29.1) 199 (25.7)
216 (21.1) 233 (22.7) 302 (29.4) 275 (26.8)
187 (24.1) 289 (37.2) 203 (26.2) 97 (12.5)
231 (22.5) 380 (37.0) 290 (28.3) 125 (12.2)
188 (24.2) 460 (59.3) 121 (15.6) 7 (0.9)
255 (24.9) 607 (59.1) 153 (14.9) 11 (1.1)
83 (10.7) 333 (42.9) 211 (27.2) 87 (11.2) 60 (7.7) 2 (0.3)
112 (10.9) 433 (42.2) 282 (27.5) 120 (11.7) 76 (7.4) 3 (0.3)
316 (40.7) 460 (59.3)
436 (42.5) 590 (57.5)
414 (53.4) 320 (41.2) 42 (5.4)
424 (41.3) 377 (37.0) 225 (21.9)
52 (6.7) 104 (13.4) 124 (16.0) 165 (21.3) 174 (22.4) 157 (20.2)
59 (5.7) 128 (12.5) 166 (16.2) 231 (22.5) 234 (22.8) 208 (20.3)
507 (65.3) 264 (34.1) 5 (0.6)
703 (68.4) 317 (31.0) 6 (0.6)
641 (82.6) 133 (17.1) 2 (0.3)
845 (82.3) 178 (17.4) 3 (0.3)
Y. Sellier et al. / Environment International 66 (2014) 165–173
university hospitals, France (Slama et al., 2009). Conception dates ranged from 2002 to 2005 in both cities. Offspring birth weight was obtained from maternity records. Personal data (Table 1) were collected by questionnaires during pregnancy and after birth. We excluded women who did not speak or write French, had personal history of diabetes, multiple pregnancy, or intention to deliver outside the university hospital or to move out of the study area within 3 years. The study was approved by all relevant ethical committees (Comité Consultatif pour la Protection des Personnes dans la Recherche Biomédicale, Le Kremlin-Bicêtre University Hospital, and Commission Nationale de l'Informatique et des Libertés), and all participating women gave informed written consent for their own participation and that of their children. 2.2. Exposure assessment to air pollutants and study area Exact home addresses of the women during pregnancy were collected and geocoded. We considered four exposure time-windows: the whole pregnancy, and each trimester of the pregnancy (92 days per trimester if no delivery). We calculated the exposure of women who moved during a time-window (11.8%) using time-weighted averages of levels at each home address. 2.2.1. Nearest AQMS model We obtained NO2 and PM10 hourly concentrations from the permanent AQMS (Fig. 1a, e, i, k) and considered circular buffers with a 10 km radius around each AQMS for PM10 and 5 km for NO2 (see Supplemental Material). We also studied buffers of 2 km and 1 km. AQMSs located up to within 2.5 km of the limits of the study area were included, except traffic (i.e., located b5 m from a road with traffic levels of N10,000 vehicles/day) (ADEME, 2002) and industrial AQMS. 2.2.2. Temporally-adjusted geostatistical model NO2 measurement campaigns were conducted in Nancy in 2002 and Poitiers in 2005 and repeated throughout the year so as to capture seasonal variations (see Supplemental Material). The campaigns were conducted in the metropolitan areas of both cities using Passam tubes and Palmes diffusive samplers (Palmes et al., 1976). The tubes were located so as to give measurements of air pollution levels in urban, suburban and rural areas (98 locations in Nancy, 61 locations in Poitiers). The estimated annual concentrations were smoothed over the whole area using universal kriging on a 50 × 50 m grid using Isatis Software version 6.06 (Geovariances, Fontainebleau, France) (Fig. 1b, f). These spatiallyresolved estimates were then combined to time-specific measurements from AQMSs to capture temporal variation relevant to each timewindow (Lepeule et al., 2010). 2.2.3. Land-use regression model From the annual NO2 averages estimated from the measurement campaigns described above, we built a LUR model following the methodology of the ESCAPE (European Study of Cohorts for Air Pollution Effects) project and reviewed by its exposure working group (Fig. 1d, h) (Beelen R et al., unpublished data) (see Supplemental Material). From the 98 locations measured in Nancy and 61 measured in Poitiers, respectively, 39 and 7 locations were discarded to meet the requirements of the ESCAPE protocol, and 1 and 3 locations were discarded because they were influential observations (according to Cook's D value), leaving 58 locations in Nancy and 51 in Poitiers. The LUR equations included the following concentration predictors: total traffic load of major roads (1 km), length of major roads (25 m), and length of all roads (50 m) in Nancy; total traffic load of all roads (within 25 m), number of inhabitants (300 m), presence of industry (300 m), and length of major roads (1 km) in Poitiers. Cross-validation R2 was 0.67 in Nancy and 0.80 in Poitiers. These equations were applied to each woman's address. As for the TAG model, these spatially-resolved estimates were combined to AQMS time-specific measurements.
167
2.2.4. Dispersion model NO2 and PM10 hourly concentrations were modeled at women's home addresses using the quasi-Gaussian Atmospheric Dispersion Modeling System (ADMS) Urban 3.1 (Cambridge Environmental Research Consultants, Cambridge, United Kingdom) (AIRLOR, 2010; Galineau et al., 2011) (see Supplemental Material). Pearson correlations between the hourly model predictions and the measures at the AQMS locations ranged from 0.50 to 0.65 for NO2 in Nancy and 0.58 to 0.62 for PM10. In Poitiers, the correlations ranged from 0.65 to 0.70 for NO2 and 0.85 to 0.95 for PM10. Model outputs for annual mean concentrations in NO2 and PM10 in 2005 were predicted (Isatis Software, Géovariances, Fontainebleau) using a regular grid with points every 300 m and the ADMS intelligent gridding setting, which further places 2 points on either side of the road distributed every 100 m along the road. Annual concentrations were mapped on 50 × 50 m interpolation grid (Fig. 1c, g, j, l). 2.2.5. Study area For each pollutant, the study area was the geographic area at the intersection between all exposure models (Fig. 1). The study area included urban (city), suburban (metropolitan area according to the regional authorities' classification of communities) and rural areas. 2.3. Statistical analysis We performed variance decomposition of the PM10 nearest AQMS and NO2 LUR models (not possible for the dispersion model) to assess their spatial and temporal components (Lepeule et al., 2010). For each pair of exposure models, we compared exposure estimates using the Wilcoxon signed rank test, Spearman rank correlation (r), and Bland– Altman plots adapted for more than two measures when there is no gold standard (Seed, 2000). We calculated Spearman correlation by city, type of area, buffer size and exposure time-window. The relationship between exposure estimated by each exposure model and birth weight was estimated by separate linear regressions adjusted for the following a priori-chosen covariates: gestational age (linear and quadratic), sex of the newborn, maternal height, pre-pregnancy weight (broken stick regression with a knot at 60 kg), birth order, maternal age at the end of education, city, month of conception, and active and passive smoking during the relevant time-window under study. We examined effect modification by city by including an interaction term with the exposure variable. The exposure–response relationships were further assessed for each buffer size. Time-windows of exposure were first examined separately. We then simultaneously adjusted for the three trimesters of pregnancy. Results were considered statistically significant when p-value b 0.05. Statistical analyses were carried out with SAS version 9.2 (SAS Institute, Cary, NC). 3. Results 3.1. Study area and study sample The PM10 study area included 1026 women and the NO2 study area included 776 women. Women mainly lived in urban and suburban settings (Table 1). For most characteristics, mothers living within different distances from a monitor did not clearly differ (see Supplemental Material, Tables S1 and S2). 3.2. Comparison of exposure estimates according to the exposure model As results by trimester of the pregnancy for both pollutants were similar to those for the whole pregnancy, we only describe the latter. Depending on the exposure model, mean pregnancy concentrations were between 22 and 29 μg/m3 for NO2 and between 20 and 21 μg/m3 for PM10. All exposure models showed lower exposure levels in rural than in urban and suburban areas (Fig. 2) and lower averages in Poitiers
168
NO2
b) POITIERS : Geostatistical model
e) NANCY : Nearest AQMS model
f) NANCY : Geostatistical model
c) POITIERS : Dispersion model
d) POITIERS : Land-use regression
g) NANCY : Dispersion model
h) NANCY : Land-use regression
i) POITIERS : Nearest AQMS model
j) POITIERS : Dispersion model
PM10
k) NANCY : Nearest AQMS model
l) NANCY : Dispersion model
Fig. 1. NO2 (a–h) and PM10 (i–l) concentration levels in the metropolitan area of Poitiers and Nancy in 2005 estimated by four models: a, e, i, k) nearest AQMS (air quality monitoring station) model; b, f) geostatistical model; c, g, j, l) dispersion model; and d, h) land-use regression model.
Y. Sellier et al. / Environment International 66 (2014) 165–173
a) POITIERS : Nearest AQMS model
Y. Sellier et al. / Environment International 66 (2014) 165–173
169
32.3±10 24.9±5 28.1±12 23.0±8
25.5±9 23.0±4 25.8±10 21.1±6
16.7±3 17.9±3 16.3±4 12.3±4
20.8±3
20.7±3
20.3±4
21.1±4
20.4±5
20.5±3
20.8±4
19.1±4
35
28.6±10 23.7±5 26.5±11 21.7±7
0
0
5
10
10
20
15
30
20
25
40
30
50
60
AQMS TAG LUR Dispersion
776 770 776 776
Whole area
414 412 414 414
320 319 320 320
Urban area
Suburban area
42
42
39
42
1025
Rural area
1026
Whole area
424
424
Urban area
NO2
377
377
224
Suburban area
n
225
Rural area
PM10
area
Pollutant
Fig. 2. Box plots and average ± SD of maternal pregnancy exposure to NO2 and PM10 (μg/m3) estimated by four exposure models, EDEN mother–child cohort. Boxes represent the 25th, 50th, and 75th percentiles. Whiskers represent the lower quartile–1.5 interquartile range and the upper quartile + 1.5 interquartile range.
compared to Nancy (see Supplemental Material, Fig. S1). For both pollutants, all pairwise comparisons of concentration distributions differed between exposure models (p-values b 0.001).
correlation among exposure models other than the nearest AQMS was not affected by the buffer size. When both cities were considered separately, the correlations with the nearest AQMS were much lower in Poitiers compared to Nancy (see Supplemental Material, Table S3). Compared to other models, the nearest AQMS tended to underestimate NO2 exposure when the concentration was low and to overestimate when the concentration was high (see Supplemental Material, Fig. S2). To a lesser extent, we observed the opposite trend for the TAG model. Since the LUR was the only model to capture exposure estimates N45 μg/m3 (Fig. 2), the scatter plot indicated an overestimation from this model for high concentrations (see Supplemental Material, Fig. S2). The dispersion model seemed generally to underestimate the exposure levels.
3.2.1. Nitrogen dioxide Pregnancy exposure estimates were subject to stronger spatial variations for the seasonalized LUR (accounting for 99% of the variance of exposure) than for the nearest AQMS (95%) and TAG (92%) models. The spatial contribution was lower when we considered trimesterspecific windows with 76–87% for the LUR, 78–82% for the nearest AQMS and 47–61% for the TAG. As for the whole area, the range of exposure estimates was largest for the LUR and dispersion models, and lowest for the TAG model (Fig. 2). Mean exposure decreased from the nearest AQMS, then the LUR, the TAG and finally to the dispersion models (for the whole area of the study). Correlations between model exposure estimates were high (≥0.70), except for the nearest AQMS with both the LUR (r = 0.54) and dispersion models (r = 0.63, Table 2). The highest correlation was between the LUR and dispersion with r = 0.87. Nearest AQMS better correlated with other models when we restricted the population to women living within 2 km to 1 km from the nearest AQMS (Table 2). As expected, the
3.2.2. Particles The spatial component of the nearest AQMS accounted for 22% of the variation of PM 10 exposure during pregnancy and for 15– 19% by trimester of the pregnancy. The mean exposure level produced by the nearest AQMS and dispersion models was similar, although the dispersion model exhibited a larger range (Fig. 2). The correlation between the nearest AQMS and the dispersion model
Table 2 Correlation between pairs of four exposure models estimating NO2 and PM10 levels during pregnancy, EDEN mother–child cohort. Pollutant
Exposure models
10 n NO2
AQMS-
TAG-
PM10
LURAQMS-
Type of areaa
Buffer size of the nearest AQMS (km)
TAG LUR Dispersion LUR Dispersion Dispersion Dispersion
1025
5
2
1
Urban
Suburban
Rural
r
n
r
n
r
n
r
n
r
n
r
n
r
0.81
770 776 776 770 770 776 790
0.70 0.54 0.63 0.79 0.79 0.87 0.81
428 431 431 428 428 431 372
0.76 0.59 0.65 0.78 0.75 0.86 0.85
157 158 158 157 157 158 136
0.85 0.76 0.71 0.77 0.63 0.84 0.85
412 414 414 412 412 414 424
0.73 0.54 0.62 0.78 0.75 0.86 0.86
319 320 320 319 319 320 377
0.51 0.46 0.53 0.73 0.75 0.84 0.75
39 42 42 39 39 42 224
0.66 0.59 0.69 0.87 0.73 0.77 0.83
AQMS: air-quality monitoring station model (home's nearest AQMS in a 5 km buffer for NO2 and 10 km buffer for PM10), TAG: temporally adjusted geostatistical model, LUR: land use regression model, n: sample size, r: Spearman correlation. a Using a 5 km buffer around each AQMS for NO2 and a 10 km buffer for PM10.
170
Y. Sellier et al. / Environment International 66 (2014) 165–173
each 10 μg/m3 increase in PM10 exposure as estimated by the dispersion model during the 1st trimester of pregnancy, and a 175 g (95%CI: 48; 301) increase in birth weight associated with each 10 μg/m3 increase in PM10 exposure during the 2nd trimester of pregnancy.
was 0.81, a value higher than for NO2. The correlation improved a little between 5 km and 2 km (Table 2) and was higher in Poitiers (r = 0.73) than in Nancy (r = 0.44) (see Supplemental Material, Table S3). The nearest AQMS tended to provide higher PM10 concentration both at low and high levels (see Supplemental Material, Fig. S2).
4. Discussion 3.3. Effect estimates of NO2 and PM10 on birth weight Fetal life is a sensitive period during which exposure to toxicants may have immediate and also delayed adverse effects on health (Perera and Herbstman, 2011). Because timing of exposure is likely to be a key feature for explaining the health effects of fetal exposure to toxicants, both spatial and temporal exposure variability require to be captured by exposure metrics. We developed four exposure models with different spatial and temporal resolution, and compared their exposure estimates for pregnant women and effect estimates on infant birth weight. For PM10, the nearest AQMS and dispersion models indicated consistent results both in terms of exposure estimates and association with birth weight. For NO2, although there was an overall moderate to good correlation among the four exposure models (r in the 0.54–0.87 range), the direction of the association with birth weight differed according to the model used for estimating air pollution exposure.
NO2 effect estimates on birth weight generally had large confidence intervals centered around zero and the sign of the point estimate varied according to the exposure model (Fig. 3). As for the exposure timewindow, considering results from the 4 exposure models, there was no clear pattern of sensitivity to NO2 through the pregnancy. PM10 levels were positively associated with birth weight, with significant associations during the second trimester of pregnancy as estimated by the dispersion model. For both pollutants, we did not find any effect modification by city. As for uncertainty, for both pollutants, health effect estimates based on the dispersion and TAG models had the largest variance, while those from the AQMS and LUR models provided the lowest variance. When restricting to smaller buffer size around AQMS, the effect estimates of NO2 on birth weight decreased for all exposure models and all time-windows. We observed the opposite trend for PM10. Further adjusting for all exposure time-windows (correlations ≤ 0.86, see Supplemental Material, Table S4) changed subsequently the health effect estimates (see Supplemental Material, Fig. S3). For all exposure models and for both pollutants, the association with birth weight tended to be negative with exposure during the 1st trimester of pregnancy, positive with the 2nd trimester of pregnancy and null with the 3rd trimester of pregnancy. These associations were stronger for PM10 than for NO2, with a 90 g (95%CI: −220; 42) decrease in birth weight associated with
4.1. Characteristics and agreement of exposure models For NO2, we observed good correlations (i.e. ≥ 0.79) between the TAG, LUR and dispersion models, while the nearest AQMS exhibited moderate correlations from 0.54 to 0.70 with the aforementioned models. Geostatistical models are generally more spatially resolved than AQMS-based models, but do not capture fine within-urban variations in air pollution due to local sources. Therefore, geostatistical and
NO2 AQMS TAG LUR DISPERSION
Change in birth weight in gr
200 150 100 50 0 -50 -100 -150 -200 -250
5km
2km
1km
5km
2km
1km
5km
2km
1km
5km
2km
(704)
(414)
(164)
(703)
(405)
(154)
(704)
(398)
(147)
(706)
(395)
1st trimester
2nd trimester
3rd trimester
Pregnancy
1km Buffer size (147) (n) Exposure window
PM10 Change in birth weight in gr
500
AQMS
DISPERSION
400 300 200 100 0 -100 -200 -300 -400
10km
5km
2km
1km
10km
5km
2km
1km
10km
5km
2km
1km
10km
5km
2km
(926)
(730)
(364)
(147)
(890)
(688)
(337)
(132)
(891)
(692)
(332)
(126)
(940)
(724)
(341)
1st trimester
2nd trimester
3rd trimester
Pregnancy
1km Buffer size (126) (n) Exposure window
Fig. 3. Change in mean birth weight (g) for a 10 μg/m3 increase in air pollutant concentration, as a function of the size of the buffer considered around each AQMS, adjusted for factors as described in “Methods” section. Error bars indicate 95%CIs.
Y. Sellier et al. / Environment International 66 (2014) 165–173
AQMS models are usually thought to represent regional air quality. Accordingly, exposure estimates from our TAG model exhibited the lowest variability (range and SD) of the four exposure models. LURs allow capturing within-urban variability in air pollutant concentrations, which is illustrated by the stronger contribution of the spatial component in the LUR compared to the TAG model. Dispersion models simulate the chemical reactions, transport and dispersion of pollutants and can capture fine contrasts in both spatial and temporal scales thanks to appropriate input data. In our study, the TAG model was well correlated with both the LUR and dispersion models (r = 0.79), which confirms its ability to capture some of the spatial variability in NO2 concentrations. The LUR and dispersion models, which were highly correlated (r = 0.86), provided highly spatially-resolved estimates (Fig. 1). 4.2. Influence of the location of AQMS and sources of air pollution Exposure estimates from AQMS-based models strongly depend on the location and density of monitors, as previously extensively discussed (e.g., Lepeule et al., 2010). The somewhat different ranges and correlations of NO2 levels between the nearest AQMS and other models in Poitiers (see Supplemental Material, Table S3, Fig. S1) are likely due to the influence of the city-center AQMS, which experienced very high NO2 concentrations (Lepeule et al., 2010, Fig. 2, station M). More generally, the stronger improvement of the correlation between the nearest AQMS and the other models for NO2 from 2 to 1 km (6– 16 units, Table 2) compared to 5 to 2 km (2–6 units) also reflects the fact that AQMS is usually only representative of air pollution levels in a very small surrounding area (Lebret et al., 2000). As for PM10, the correlation between the nearest AQMS and dispersion models was not substantially influenced by the buffer size around the monitors. These results likely own to the quite homogenous distribution of PM10 concentrations across the study area, which is illustrated by the low spatial variability of the nearest AQMS model and is in line with the results obtained in European countries (Gulliver et al., 2011). When the buffer size of the AQMS model decreased, effect estimates moved usually further from the null (except for the PM10 1 km buffer). These results are unlikely to be due to a selection bias (see Supplementary Materials, Tables S1 and S2) but rather result from a decrease in home address exposure misclassification. PM10 in Nancy and Poitiers originate from both regional pollution episodes and local sources such as traffic and residential emissions. Pollutant concentrations were higher in Nancy, which is more urbanized and densely populated compared to Poitiers. In both cities, NO2 concentrations were higher in urban and suburban areas, where traffic, urban heating and industries are the main sources, compared to rural areas.
171
ours. In our study, the LUR model predicted the highest concentrations, especially in Nancy where exposures between 45 and 73 μg/m3 were estimated during pregnancy. This result is in agreement with those of Marshall et al. (2008) and the levels estimated in Nancy are compatible with its urbanization and the measures of monitors performed in traffic proximity. In the Vancouver study, the authors reported significant positive associations between NO2 exposure during pregnancy and term low birth weight (b2500 g) when they used an AQMS-based approach, but no association when they used the LUR (Brauer et al., 2008). Although we did not find any significant association of continuous birth weight with whole pregnancy exposure, the pattern of our results (stronger effects with the nearest AQMS than with the LUR) is compatible with this study. Wu et al. (2011) estimated NO2 during pregnancy and found a correlation of 0.57 between a temporally-adjusted LUR and the nearest AQMS. They reported a positive association of preterm birth with the nearest AQMS exposure, but a negative association with the LUR estimate. In that same study, Laurent et al. reported positive associations between all pollutants (except ozone) measured by the nearest AQMS, a LUR or a dispersion model and term birth weight (Laurent et al., 2013). In a study estimating air pollution exposure at home and workplace during pregnancy, Madsen et al. (2010) reported a very low correlation between exposures as estimated by a fine spatial scale dispersion model and the fixed monitor of the city (0.07 for NO2 and 0.19 for PM10) and found negative associations with birth weight for both pollutants as estimated by the monitor, but positive associations using the dispersion model. Although we do not have any biological plausible explanation for these positive associations with birth weight, our results agree with those studies. However, such results are not corroborated by a recent meta-analysis (Dadvand et al., 2013) indicating adverse effects of whole pregnancy exposure to particles on term birth weight. In comparison, the lack of significance of our results regarding whole pregnancy NO2 and birth weight may be due to our limited statistical power. It might also be viewed as a consistent result among exposure models indicating no significant association with continuous birth weight, perhaps because NO2 is not a relevant marker for air pollution effects on birth weight. In our sensitivity analyses adjusted for all trimesters of exposure, we saw, for all exposure models and both pollutants, a trend for a negative association of birth weight with exposure during the first trimester of pregnancy, while there was a trend for a positive association with the second trimester of pregnancy. Such varying effects of exposure during the different trimesters of pregnancy might be artifacts due to the use of attained weight at birth, which does not represent the patterns of fetal growth along the three trimesters. A way to overcome this issue would be to use ultrasound measurements during the course of the pregnancy (Aguilera et al., 2010; van den Hooven et al., 2012).
4.3. Comparison with the epidemiological literature 4.4. Exposure assessment and time–activity patterns Few papers have compared different methods for estimating air pollution exposure in terms of impact on health effect estimates (Laurent et al., 2013; Sahsuvaroglu et al., 2009; Son et al., 2010). As reproductive epidemiology often deals with exposure windows of 9 months or less, both spatial and temporal resolution exposure contrasts should be considered. Our results regarding the nearest AQMS and TAG NO2 models have been previously discussed (Lepeule et al., 2010). We therefore focus here on comparisons involving the fine spatial scale LUR and dispersion models as well as PM10. In a term-born infant study in California, Laurent et al. found NO2 correlations of 0.57 between the nearest AQMS and LUR (Laurent et al., 2013). In a comparison study of annual NO2 levels in Vancouver, Marshall et al. (2008) found correlations of 0.54 between the nearest AQMS and LUR, 0.61 between the nearest AQMS and dispersion models and 0.49 between the LUR and dispersion models, which are comparable to our results (0.54, 0.63, and 0.87 respectively). They used a larger 10 km buffer around monitors and their dispersion model accounted for downtown versus suburban variations only, which likely explains its low correlation with the LUR compared to
Another explanation for the inconsistent relationships with birth weight among the four exposure models and for the unexpected positive relationships lies in the exposure measurement error. As in most epidemiological studies, we considered air pollution exposure of women at their home address as a surrogate of the personal exposure, which includes indoor and outdoor sources and transportations, as a function of time spent in these microenvironments. In the Vancouver study, Nethery et al. (2008) evaluated LUR estimates as predictors of personal exposure during pregnancy. They reported that accounting for work outdoor concentrations improved LUR estimates, whose correlation with personal exposure was 0.28, compared to 0.05 when the LUR model focused on home address only. However, these results suggest that there is ample room for improvements of air pollution exposure estimates. For example, time spent commuting was shown to be responsible for 29% of personal benzene exposure (Bruinen de Bruin et al., 2008). Highly spatially resolved air pollution models such as our LUR and dispersion models predict levels that are very specific to the
172
Y. Sellier et al. / Environment International 66 (2014) 165–173
location, usually the subject's home address, while exposure models relying on monitors or interpolation methods provide a mean level over a broader area, which may include other microenvironments of the subject. Therefore, not accounting for time–activity patterns when using highly spatially-resolved exposure models may create a similar or bigger exposure misclassification than, for example, the error made by the nearest AQMS model, and therefore produce biased associations with birth weight.
FRM, Inserm, IReSP, Nestlé, French Ministry of Health, ANR, Univ. Paris-Sud, InVS, ANSES and MGEN. The Team of Environmental Epidemiology at Inserm U823 is supported by an AVENIR/ATIP grant. The authors declare that they have no competing financial interests. Appendix A. Supplementary data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.envint.2014.02.001.
4.5. Generalizability of the results and relevance of exposure models References A recent study comparing four PM10 exposure models in London indicated poor performance of the nearest AQMS and kriging (both based on 52 monitors) to predict long-term ambient PM and slightly better performances of the LUR and dispersion models (Gulliver et al., 2011). The authors concluded that these models rather predict the average level of the urban area than the local patterns of variation and that LUR might be the most robust method to predict long-term exposure in complex areas. However, the generalizability of results from exposure model comparison is limited to areas with similar characteristics, such as regional and local sources of emissions, monitor coverage, and weather and topography of the area. Furthermore, the relevance of the method used to assess exposure depends on the spatio-temporal variability of the pollutants, the time-window length and endpoints considered and the mobility of the population. For example, assessment of exposure to PM10, due to its usually low spatial variability, may not be much improved by the use of highly-spatially resolved methods such as our dispersion model, which is computationally intensive and requires data that may not be easily available everywhere. The LUR is thought to perform at least as well as a dispersion model when a sufficient dense network of monitors is used (Dijkema et al., 2011). In ESCAPE, LUR models relying on 40 NO2 measurement sites have been found efficient as long as the measurement sites included a large range of potential sources of air pollution variability (Beelen R et al., unpublished data). Other studies consider that a larger number is required (Basagana et al., 2012). Our LUR included at least 51 sites in each city distributed so as to represent such variability. However, one should bear in mind that LUR is not initially designed for reflecting temporal variations and may not account for the meteorology, which might be an issue for certain areas. Fixed monitor data are easily accessible, usually free and provide high temporally resolved models with however a usually poor spatial resolution, which has led to the conclusion that these exposure models are likely to introduce exposure error and bias the effect estimates. However, studies such as short-term health effects might still benefit from monitoring data in certain conditions (Sarnat et al., 2010). 5. Conclusion Our study adds to the growing body of literature on maternal air pollution exposure effects on fetal growth by showing that different spatiotemporal approaches of air pollution exposure assessment, although with well correlated predictions, provide different rankings of the subjects regarding their exposure and may lead to different conclusions regarding associations with health endpoints. Accounting for time– activity patterns of the subjects rather than solely considering the home addresses might be a key issue for highly-spatially resolved exposure models compared to models based on between-community comparisons. Acknowledgments We are indebted to the midwife research assistants (L. Douhaud, S. Bedel, B. Lortholary, S. Gabriel, M. Rogeon, and M. Malinbaum) for data collection and to P. Lavoine for checking, coding, and data entry. This work was supported by ANSES. The EDEN cohort is funded by
ADEME. Classification and criteria for setting up air-quality monitoring stations. Paris ADEME ÉditionsParis ADEME Éditions; 2002. Aguilera I, Garcia-Esteban R, Iniguez C, Nieuwenhuijsen MJ, Rodriguez A, Paez M, et al. Prenatal exposure to traffic-related air pollution and ultrasound measures of fetal growth in the INMA Sabadell cohort. Environ Health Perspect 2010;118:705–11. AIRLOR. Concentration en polluants dans l'air ambiant extérieur aux domiciles des sujets de la cohorte épidémiologique EDEN. AIRLORNancy: AIRLOR; 2010. Basagana X, Rivera M, Aguilera I, Agis D, Bouso L, Elosua R, et al. Effect of the number of measurement sites on land use regression models in estimating local air pollution. Atmos Environ 2012;54:634–42. Brauer M, Lencar C, Tamburic L, Koehoorn M, Demers P, Karr C. A cohort study of traffic-related air pollution impacts on birth outcomes. Environ Health Perspect 2008;116:680–6. Briggs D. The role of GIS: coping with space (and time) in air pollution exposure assessment. J Toxicol Environ Health A 2005;68:1243–61. Bruinen de Bruin Y, Koistinen K, Kephalopoulos S, Geiss O, Tirendi S, Kotzias D. Characterisation of urban inhalation exposures to benzene, formaldehyde and acetaldehyde in the European Union: comparison of measured and modelled exposure data. Environ Sci Pollut Res Int 2008;15:417–30. Cyrys J, Hochadel M, Gehring U, Hoek G, Diegmann V, Brunekreef B, et al. GIS-based estimation of exposure to particulate matter and NO2 in an urban area: stochastic versus dispersion modeling. Environ Health Perspect 2005;113:987–92. Dadvand P, Parker J, Bell ML, Bonzini M, Brauer M, Darrow LA, et al. Maternal exposure to particulate air pollution and term birth weight: a multi-country evaluation of effect and heterogeneity. Environ Health Perspect 2013;121:267–373. Dijkema MB, Gehring U, van Strien RT, van der Zee SC, Fischer P, Hoek G, et al. A comparison of different approaches to estimate small-scale spatial variation in outdoor NO(2) concentrations. Environ Health Perspect 2011;119:670–5. Galineau J, Hulin A, Caini F, Marquis N, Lepeule J, Slama R. Estimation of exposure to urban air pollution in two cities using a Gaussian dispersion model: the EDEN-Air project. Abstracts of the 23rd Annual Conference of the International Society of Environmental Epidemiology (ISEE) September 13–16, 2011, Barcelona, SpainEnviron Health Perspect; 2011. Gulliver J, de Hoogh K, Fecht D, Vienneau D, Briggs D. Comparative assessment of GIS-based methods and metrics for estimating long-term exposures to air pollution. Atmos Environ 2011;45:7072–80. Jerrett M, Burnett RT, Ma R, Pope III CA, Krewski D, Newbold KB, et al. Spatial analysis of air pollution and mortality in Los Angeles. Epidemiology 2005;16:727–36. Laurent O, Wu J, Li L, Chung J, Bartell S. Investigating the association between birth weight and complementary air pollution metrics: a cohort study. Environ Health 2013;12:18. Lebret E, Briggs D, van Reeuwijk H, Fischer P, Smallbone K, Harssema H, et al. Small area variations in ambient NO2 concentrations in four European areas. Atmos Environ 2000;34:177–85. Lepeule J, Caini F, Bottagisi S, Galineau J, Hulin A, Marquis N, et al. Maternal exposure to nitrogen dioxide during pregnancy and offspring birth weight: comparison of two exposure models. Environ Health Perspect 2010;118:1483–9. Madsen C, Gehring U, Walker SE, Brunekreef B, Stigum H, Naess O, et al. Ambient air pollution exposure, residential mobility and term birth weight in Oslo, Norway. Environ Res 2010;110:363–71. Marshall JD, Nethery E, Brauer M. Within-urban variability in ambient air pollution: comparison of estimation methods. Atmos Environ 2008;42:1359–69. Nethery E, Leckie SE, Teschke K, Brauer M. From measures to models: an evaluation of air pollution exposure assessment for epidemiological studies of pregnant women. Occup Environ Med 2008;65:579–86. Palmes ED, Gunnison AF, DiMattio J, Tomczyk C. Personal sampler for nitrogen dioxide. Am Ind Hyg Assoc J 1976;37:570–7. Pereira G, Nassar N, Cook A, Bower C. Traffic emissions are associated with reduced fetal growth in areas of Perth, Western Australia: an application of the AusRoads dispersion model. Aust N Z J Public Health 2011;35:451–8. Perera F, Herbstman J. Prenatal environmental exposures, epigenetics, and disease. Reprod Toxicol 2011;31:363–73. Pouliou T, Kanaroglou PS, Elliott SJ, Pengelly LD. Assessing the health impacts of air pollution: a re-analysis of the Hamilton children's cohort data using a spatial analytic approach. Int J Environ Health Res 2008;18:17–35. Rahmalia A, Giorgis-Allemand L, Lepeule J, Philippat C, Galineau J, Hulin A, et al. Pregnancy exposure to atmospheric pollutants and placental weight: an approach relying on a dispersion model. Environ Int 2012;48:47–55. Rosenlund M, Forastiere F, Stafoggia M, Porta D, Perucci M, Ranzi A, et al. Comparison of regression models with land-use and emissions data to predict the spatial distribution of traffic-related air pollution in Rome. J Expo Sci Environ Epidemiol 2008;18: 192–9.
Y. Sellier et al. / Environment International 66 (2014) 165–173 Sahsuvaroglu T, Jerrett M, Sears MR, McConnell R, Finkelstein N, Arain A, et al. Spatial analysis of air pollution and childhood asthma in Hamilton, Canada: comparing exposure methods in sensitive subgroups. Environ Health 2009;8:14. Sarnat SE, Klein M, Sarnat JA, Flanders WD, Waller LA, Mulholland JA, et al. An examination of exposure measurement error from air pollutant spatial variability in time-series studies. J Expo Sci Environ Epidemiol 2010;20:135–46. Seed P. sbe33: comparing several methods of measuring the same quantity. Stata Tech Bull 2000;55:2–9. Slama R, Morgenstern V, Cyrys J, Zutavern A, Herbarth O, Wichmann HE, et al. Traffic-related atmospheric pollutants levels during pregnancy and offspring's term birth weight: a study relying on a land-use regression exposure model. Environ Health Perspect 2007;115:1283–92.
173
Slama R, Thiebaugeorges O, Goua V, Aussel L, Sacco P, Bohet A, et al. Maternal personal exposure to airborne benzene and intrauterine growth. Environ Health Perspect 2009;117:1313–21. Son JY, Bell ML, Lee JT. Individual exposure to air pollution and lung function in Korea: spatial analysis using multiple exposure approaches. Environ Res 2010;110:739–49. van den Hooven EH, Pierik FH, de Kluizenaar Y, Willemsen SP, Hofman A, van Ratingen SW, et al. Air pollution exposure during pregnancy, ultrasound measures of fetal growth, and adverse birth outcomes: a prospective cohort study. Environ Health Perspect 2012;120:150–6. Wu J, Wilhelm M, Chung J, Ritz B. Comparing exposure assessment methods for traffic-related air pollution in an adverse pregnancy outcome study. Environ Res 2011;111:685–92.
